We are currently in the midst of Earth's sixth extinction event, and measuring biodiversity trends in space and time is essential for prioritizing limited resources for conservation. At the same time, the scope of the necessary biodiversity monitoring is overwhelming funding for professional scientific monitoring. In response, scientists are increasingly using citizen science data to monitor biodiversity. But citizen science data are 'noisy', with redundancies and gaps arising from unstructured human behaviours in space and time. We ask whether the information content of these data can be maximized for the express purpose of trend estimation. We develop and execute a novel framework which assigns every citizen science sampling event a marginal value, derived from the importance of an observation to our understanding of overall population trends. We then make this framework predictive, estimating the expected marginal value of future biodiversity observations. We find that past observations are useful in forecasting where high-value observations will occur in the future. Interestingly, we find high value in both 'hotspots', which are frequently sampled locations, and 'coldspots', which are areas far from recent sampling, suggesting that an optimal sampling regime balances 'hotspot' sampling with a spread across the landscape.
Keywords: biodiversity; citizen science; dynamic models; predictive modelling; spatial and temporal sampling.